Data center cooling
Almost every watt delivered to IT equipment comes back out as heat, which makes cooling capacity a direct function of electrical load. Where air stops being enough, and what the liquid loop adds.
Cooling is an electrical problem wearing mechanical clothes
Essentially every watt delivered to IT equipment comes back out as heat. That makes cooling capacity a direct function of electrical load, and it is why cooling and power sizing cannot be done independently — a rack density decision made on the electrical side lands immediately on the mechanical side.
The genuine shift over the last few years is that air alone has stopped being sufficient at the top end. Somewhere above roughly 30–40 kW per rack, moving enough air becomes impractical, and heat has to be picked up in liquid closer to the chip. That does not eliminate air — it adds a second loop, with its own equipment, its own failure modes and its own procurement lead time.
| Family | Typical configuration | Indicative lead time | Standards |
|---|---|---|---|
| Coolant Distribution Units (CDU)The interface between a facility water loop and the technology loop that reaches the servers. | 300 kW – 1.4 MW Liquid-liquid or liquid-air N or N+1 |
~20 weeks | ASHRAE UL 60335 |
| CRAH / CRAC UnitsRoom-level air cooling. CRAH uses chilled water; CRAC has its own refrigeration circuit. | 30–300 kW Chilled water or DX Upflow or downflow |
~18 weeks | ASHRAE 90.4 AHRI |
| Rear-Door Heat ExchangersCaptures heat at the rack before it reaches the room. Passive versions add no fan power. | 35–100 kW per rack Passive or active 600 / 750 mm |
~14 weeks | ASHRAE |
| Dry Coolers & ChillersWhere the heat finally leaves the site. Adiabatic units trade water consumption for capacity on hot days. | 40–600 tons Air-cooled chiller / dry cooler / adiabatic R-134a / R-513A / R-1234ze |
~24 weeks | AHRI 550/590 ASHRAE |
Specifications that decide the selection
Approach temperature
How close the supply gets to the source temperature. A tighter approach means more heat exchanger surface, more cost, and usually better annual efficiency.
Rack density, not room average
Average density hides the problem. A room averaging 8 kW per rack with a 40 kW row still needs a solution for that row.
Water quality & loop separation
The technology loop reaching servers has far tighter chemistry requirements than facility water. That separation is precisely what a CDU exists to provide.
Redundancy at the right layer
N+1 CDUs feeding a single non-redundant chilled water loop is not N+1 cooling. Redundancy has to be traced end to end, like an electrical path.
Selection criteria in practice
- Start from the electrical load, not a target PUE. Cooling capacity follows IT load almost one for one. Efficiency targets are an outcome of the design, not an input to sizing.
- Establish whether liquid is required now or later. Retrofitting a technology loop into a live room is materially harder than provisioning the pipework during construction, even if the CDUs come later.
- Check the free-cooling hours for the actual site. Economiser hours vary enormously by climate and drive the operating cost far more than nameplate efficiency does.
- Confirm water availability and discharge rules. Adiabatic and evaporative equipment buys capacity on design days by consuming water, which is increasingly constrained or reportable.
- Do not neglect the blanking panels. An open rack unit lets hot exhaust recirculate to the intake above it. It is the cheapest thermal fix available and the one most often skipped.
Typical applications
A conventional air-cooled hall uses CRAH units on a chilled water loop fed by chillers or dry coolers. As density rises, rear-door heat exchangers intercept heat at the rack without committing to full liquid cooling. High-density AI halls add a technology loop through CDUs feeding direct-to-chip cold plates, with air cooling retained for everything the liquid loop does not reach — power supplies, networking and storage still need it.
Work the numbers before you specify
Related reading
Common questions
At what rack density does air cooling stop working?
There is no hard line, but somewhere around 30–40 kW per rack moving sufficient air becomes impractical — the airflow required, and the fan power to move it, stop being reasonable. Above that, heat is generally picked up in liquid closer to the source.
What does a CDU actually do?
It separates two water loops. Facility water is comparatively dirty and variable; the technology loop that reaches server cold plates needs tightly controlled chemistry, temperature and pressure. The CDU is the controlled interface between them, and it is also where redundancy in the liquid path is usually implemented.
Does cooling need to survive a utility outage?
Yes, and this is frequently under-provisioned. IT load does not stop during transfer to generator, so heat production does not either. Chilled water thermal mass buys some time, but pumps and CDUs generally need to be on backed-up power alongside the IT load itself.
Is a lower PUE always the right target?
Not necessarily. PUE is a ratio, and it can be improved by measures that raise total energy or that trade water for electricity. It is a useful comparison metric, not an objective function — optimising it directly can produce decisions that are worse overall.
Family reference pages
Each family below has its own page: how it is specified, the standards it is built to, its indicative lead time and market price band.
Specifying cooling equipment?
Every family is in the Spec Library with its configuration axes, the standards it is built to, its indicative lead time and its market price band. Take the specification to whoever supplies it — Voltfield sells nothing and is not a route to any of it.